|
< Earlier Kibitzing · PAGE 10 OF 15 ·
Later Kibitzing> |
| Dec-06-17 | | dehanne: AlphaZero is a supercomputer possessed by the spirit of Tal. |
|
| Dec-06-17 | | ColdSong: In fact Stockfish is a 3200 elo patzer.l knew it. |
|
| Dec-06-17 | | zanzibar: Has anybody actually try to read the paper?
They spent more time/space, in the Methods section, giving details of how Stockfish functioned then they did of AlphaGo. I also found this statement curious:
<5. Chess and shogi games exceeding a maximum number of steps (determined by typical game length) were terminated and assigned a drawn outcome; ...> Given they restricted the domain knowledge so severely (i.e. no eval function), it seems that all games should be played out under the 50-move draw rule. |
|
| Dec-06-17 | | dehanne: That day when Stockfish became NN. |
|
| Dec-06-17 | | SChesshevsky: I'm no grandmaster and I really didn't analyze but I took a look at AlphaOne - Stockfish game 8 on the lichess link in an earlier post. And it looked like Stockfish might have been overly passive. Like it never heard of tempo. There was a sequence where it went something like ...Qd8 ...Qc8 ...Qc7 sequentially. Didn't seem forced. It would seem that Stockfish might have been able to put up a better defense. Could a good player or someone with a good engine check it out with an opinion? Thanks. |
|
Dec-06-17
 | | WannaBe: So.... I finally got a chance to (quickly) play through the first game in the PDF's PGN. There's no way, if you told me, that after 25-30 moves when black had no center, down on pawns that white will go on to resign. ABSOLUTELY, NO, FREAKIN', WAY.
I will go through game 2 tomorrow... |
|
| Dec-07-17 | | zanzibar: Here's a page with a link to a downloadable, normalized, version of the match PGN: https://zanchess.wordpress.com/2017... It has the ten games published in the AlphaZero paper. |
|
| Dec-07-17 | | gauer: Although I'm not sure if the quote referenced at Ostrich (Computer) is an <exact one>, it might make for an interesting <Quote of the day> - (attributed to Monroe (Monty) Newborn) if it also helps spurn discussion here. Wonder if we're in the awe stage now, after Deep Blue maintained a good March vs Kasparov nearly 2 decades ago?! |
|
Dec-07-17
 | | offramp: It's almost as if <AlphaZero> is a child prodigy. This is an amazing result, even with the short time limit and the lack of tablebases. This is the start of a new era, possibly for all mankind. |
|
| Dec-07-17 | | frogbert: I was also surprised to learn about this article and AlphaZero. I have previously read the (scientific) articles about AlphaGo, but I was among the majority that believed that self-learning neural network approaches would have a hard time improving upon state of the art chess engines. However, I think there are a few flaws in the comparison: Thinking time: an important part of modern chess engines is actually «time management» - or rather the program's capability to recognize when it needs to perform (selectively) deeper searches. This is quite intrinsic to the known limitations of the basic approach of depth-first searches. Hence, imposing fixed time/depth limits on the engine isn't quite fair. Similarly, the usage of opening books and endgame tables has been standard for quite some time - and alas one hasn't spent resources trying to improve chess engines' «understanding» of opening and endgame play - where brute force-ish depth first searches have known limitations. From the viewpoint of the people behind AlphaZero, however, this is probably of little importance, as their focus is on AI-techniques - and their recent article does put question to which approach is actually better for tactical games like chess and shogi. Still, claiming that AlphaZero reached Stockfish-level after a certain amount of time, isn't quite right, when they implicitly or explicitly remove important parts of Stockfish's functionality. Anyway, I'm very impressed and I'm looking forward to further results and exploration. |
|
| Dec-07-17 | | WorstPlayerEver: <And it looked like Stockfish might have been overly passive.> <WannaBe>
Lol I think that's an understatement. I never knew Stockfish always chooses the most defensive option. |
|
Dec-07-17
| | offramp: I think it's a repetition of chess history.
When an invincible "chess machine" comes along, like Capablanca or Karpov, who everyone says is the acme of chess strength, then someone comes along to demolish that immovable object. |
|
| Dec-07-17 | | WorstPlayerEver: <offramp>
Afaik Karpov never made any suggestions regarded to *the end of chess* Besides, when a program which *trained itself in 4 hours to become what it :is: now without human chess interference*,
I'd first want to know more details about that accomplishment. Considered that there is no such app atm. In reality. Because usually such statements are added to give things a nice twist. Let's say in 99.99% that will be the case. |
|
Dec-07-17
 | | alexmagnus: Four hours to reach Stockfish level, not to become what it is (for that it needed three days). |
|
| Dec-07-17 | | WorstPlayerEver: <alexmagnus>
Well, it didn't took me so long to figure out it's interesting, but no more than that. Of course I am thrilled to see if this current afair will lead to a drastic improvement of ELO rating. As it comes to common practice. So to speak. |
|
| Dec-07-17 | | Pravitel: What is the ELO difference between complete Stockfish and Stockfish without its opening book and endgame tablebase? |
|
| Dec-07-17 | | Bluegrey: Latest SF dev is abt 35-55 elo points stronger than the SF 8 version used in the match. Opening book and EGTB will probably add another 50 elo and if they used say 128 cores and appropriate hash and set it optimally another 40 elo. I would expect the best version of SF playing with state of the art hardware to be about 100-120 ELO stronger than the SF 8 version they tested Alphazero on. So in terms of match scores they would be about even right now |
|
| Dec-07-17 | | WorstPlayerEver: Ok. Let's gather the facts. I have seen 10 games. Which simply means there are no other 90 games so far. It's like I am surrounded by parrots. The usual.
That said, I have high suspicions that these 10 games are concocted. Each time we see *Stockfish* subsequently placing its pieces in the most awkward position. After White simply played e4-e5 at some point and places its Queen *actively* on the king side. I wonder if you guys actually use S8..
The exceptions are the 2 RL's. In which both engines *coincedentally* chose for the same RL subsubvariation. |
|
| Dec-07-17 | | WorstPlayerEver: <Bluegrey>
I was not meaning you, btw, but it seems many sources have spread the news without checking its sources. That's what I meant by parrots. Brb ;) |
|
| Dec-07-17 | | frogbert: <Four hours to reach Stockfish level> <alexmagnus> But did it really do that, when the creators of AlphaZero put arbitrary limitations on their version of Stockfish? Consider:
"Elo ratings were computed from evaluation
games between different players when given <one second per move>. a) Performance of AlphaZero
in chess, compared to 2016 TCEC world-champion program Stockfish." "Figure 1 shows the performance of AlphaZero during self-play reinforcement learning, as
a function of training steps, on an Elo scale (10). In chess, AlphaZero outperformed Stockfish
after just 4 hours (300k steps);"
I.e. that was based on giving Stockfish 1 second per move - which means very shallow search depths for a brute-force depth-first engine. Similarly: "We evaluated the fully trained instances of AlphaZero against Stockfish, Elmo and the previous
version of AlphaGo Zero (trained for 3 days) in chess, shogi and Go respectively, playing
100 game matches at tournament time controls of one minute per move." Again, this means putting an arbitrary restraint on Stockfish, not allowing it to do its own time management, i.e. control its use of selective depth searches based on the complexity of the position. If Stockfish would've been given 40 minutes for 40 moves, followed by 20 minutes for 20 moves (and so on, adding new 20 move time controls until the end of the game), my belief is that Stockfish would perform better than it did in this 100 game match. |
|
Dec-07-17
 | | Domdaniel: So the French, Sicilian, Caro-Kann and Spanish are all now defunct? I'm not too surprised, apart from the inclusion of the Ruy Lopez/ Spanish on the list. I thought the Berlin wall was still a draw. If AlphaZero is right, it isn't, which means that the Petroff is the best drawing line after 1.e4 e5. Of course I will continue to play the Sicilian and French (and Caro-Kann, Pirc, Modern, Owen, Scandinavian, etc) against humans. Never mind the ultimate correctness: anything that poses problems is good. |
|
| Dec-07-17 | | WorstPlayerEver: https://en.chessbase.com/post/the-f... A vs S game 5 after move 20:
click for larger viewNow our genius produces 21. Bg5
No further info but in the tiny paragraphs one could read that this game continued: 21... f5 22. Qf4 Nc5 The latter is an obvious blunder which S8 not even considers. It plays 21. Bg5 f5 22. Qf4 hxg5 |
|
| Dec-07-17 | | SirRuthless: So AlphaGo beat a redefined and much weakened stockfish and we are supposed to conclude what? Without the large book and tablebases these engines are functionally much weaker against humans, let alone an AI like AlphaGo so why did they place arbitrary restrictions on Stockfish? I don't think this study shows much related to the stated conclusions. |
|
| Dec-07-17 | | frogbert: < So the French, Sicilian, Caro-Kann and Spanish are all now defunct> Yep! It took a self-learning neural network to debunk Kasparov's Najdorf. Or wait - did it? ;) Seriously, we can't tell from the diagrams in the paper why AlphaZero stopped playing certain openings during learning; was it based on white's or black's results in the given opening? Or neither? :) |
|
| Dec-07-17 | | WorstPlayerEver: PS so even in this completely ridiculous position (take note: usually reached by a oh so brilliant pawn sac - or two) Alpha does not reach beta. The trouble with fraudulous people is that they usually tend to overdo it. Gotta love that word. Usually :) |
|
 |
|
|
< Earlier Kibitzing · PAGE 10 OF 15 ·
Later Kibitzing> |
Kibitzing
